Richard Sutton: Pure Generative AI Cannot Do Real Science
TL;DR. Turing Award winner Richard Sutton argues that current generative AI models lack the self-evaluation capabilities needed for true scientific discovery. - Sutton asserts generative AI often produces novelty that cannot be evaluated for actual merit without external feedback loops. - He highlights systems like AlphaGo and AlphaProof as examples where integrated evaluation enables genuine AI creativity. - Scientific discovery requires generation alongside rigorous testing and selective retention of valid outcomes, a process pure generative AI misses.
- Generative AI cannot evaluate its own outputs for scientific value.
- True scientific discovery requires variation, evaluation, and selective retention.
- Integrated evaluation loops, as seen in AlphaGo, enable genuine AI creativity.
- Current generative AI often produces novel results that are lost without human or system-driven assessment.